What this automation does
The calls that matter most to a home service company arrive at the worst times: a furnace dies at 9 p.m., a water heater leaks on Saturday, a heat wave floods the phones at 8 a.m. This agent answers every call and text, books what it can, and hands off what it should not touch.
- Answer and identify. The agent answers on your number through Twilio, greets the caller as your company, announces recording, and looks up the caller ID in your field service software to find existing customers, service addresses and equipment on file.
- Triage. It asks what is going on and classifies the call: emergency, same-day repair, routine repair, maintenance, estimate for replacement or new install, or not a service request (vendor, job applicant, existing job status).
- Handle emergencies by rule. Emergency keywords trigger a scripted safety instruction and an immediate transfer to the on-call technician or dispatcher. No booking conversation happens first.
- Qualify and quote approved fees. For bookable work, it confirms address, system type and age if known, membership or maintenance agreement status, and states the approved diagnostic or trip fee.
- Book a window. It checks available arrival windows in ServiceTitan, Housecall Pro or Jobber, offers two or three options, and books the job with a clear summary for the technician.
- Confirm and follow up. The customer gets an SMS confirmation with the window and what to expect. Calls the agent could not complete become callback tasks for the office with a transcript and summary.
When it is worth building
This is worth building when calls go unanswered. If your office misses calls during lunch, after 5 p.m., on weekends or during seasonal surges, and each booked job is worth hundreds of dollars or more, even a few recovered calls a week can justify the build.
It is not worth building if your office already answers nearly every call live, or if you use a human answering service that books well into your software and you are happy with the cost. It is also worth trying the AI answering products built for the trades before going custom. They are faster to launch and may be enough for a single-location shop.
Architecture
| Layer | What we use | Why |
|---|---|---|
| Telephony | Twilio numbers and media streams | Keeps your existing number via forwarding, handles SMS, and supports live transfer to on-call staff |
| Voice pipeline | Streaming speech-to-text, LLM, and text-to-speech with low latency | Callers notice pauses longer than a second or so; every step has to stream |
| Conversation model | Claude or OpenAI’s GPT models through the API, with tool calls | Follows the booking script, asks clarifying questions and calls booking tools |
| Field service system | ServiceTitan, Housecall Pro or Jobber API, through each vendor’s developer or partner access | Customer lookup, availability, job creation and notes |
| Rules and pricing | Code-level config for emergency keywords, service area ZIPs, fees and business hours | Safety responses and prices must never vary with model output |
| Orchestration and alerts | n8n for SMS confirmations, callback tasks and daily summaries | Easy to change without touching the voice service |
| Fallback calendar | Google Calendar for shops without a field service platform | Small operators can still book windows |
How to build it
- Listen to 50 real calls. Recordings or notes from your CSRs show how customers describe problems, which questions your best CSR asks, and where calls go wrong. This becomes the script and test set.
- Write the rules sheet with the owner. Emergency keywords and the exact safety wording, service area by ZIP, business hours, which job types can be booked by the agent, fees it may state, membership benefits, and who is on call.
- Confirm field service API access. Apply to your platform’s developer or partner program and confirm you can read customers and availability and create jobs. Start this in week one.
- Build the booking tools. Customer lookup by phone, address verification against the service area, availability by job type and zone, and job creation with a structured summary. Each tool validates its inputs in code.
- Build and tune the voice loop. Pick voices, set interruption handling and silence timeouts, and test with background noise, speakerphones and callers who give their address in pieces.
- Test against recorded scenarios. Replay your 50 calls as test cases, plus emergency, angry, out-of-area and wrong-number cases. Fix failures before going live.
- Launch on overflow and after-hours first. Route only unanswered and after-hours calls to the agent for the first few weeks. Review every transcript daily, then widen coverage.
What breaks in production
- Addresses over the phone. Street names get misheard. Read the address back, confirm the ZIP, and verify it against a geocoding service before booking.
- Availability that changes mid-call. A dispatcher fills the 8–12 window while the agent is offering it. Re-check availability at the moment of booking and offer an alternative if it is gone.
- Existing-job callers. “Where is my tech?” is not a new booking. Detect open jobs for the caller and route status questions to the office or a tracking link.
- Callers who want a price for a new system. The agent must not guess. Book an estimate visit and explain why pricing needs a site visit.
- Latency spikes. Slow model or speech responses make callers talk over the agent. Monitor response time per turn and fail over to a faster model or a transfer when it degrades.
- Silent failures. If the field service API errors during job creation, the caller may think they are booked when they are not. Always confirm by SMS only after the job exists, and alert the office on any write failure.
Timeline and cost
A single-location company with one field service platform typically takes four to eight weeks, with API approval often the longest item. Lumeor builds this in the range of $15,000 to $45,000, depending on call volume, number of locations and trades, how much membership and pricing logic is involved, and the field service integration. We built voice AI assistants on Twilio and LLMs in our AI calling platform work, and voice AI vs traditional IVR explains why phone trees lose bookings. For a full breakdown of what drives price and running cost, see how much a voice AI agent costs.
How to measure success
- Answer rate on inbound calls, overall and after hours
- Jobs booked by the agent per week, and their completed revenue
- Share of agent calls that needed a callback or human correction
- Emergency calls transferred and time to transfer
- Booking-to-cancellation rate for agent-booked jobs versus CSR-booked jobs